Equivariant graph convolutional neural network for predicting tensors of atomic Born effective charges (Equivar). Pertrained models and scripts, and datasets of ab initio tensors of atomic Born effective charges of perovskites, Li3PO4, and ZrO2

This archive contains model weights, python scripts, and data for the paper "Representing Born effective charges with equivariant graph convolutional neural networks." The equivariant graph convolutional neural network model (Equivar) performs the regression of tensors of atomic Born effective charges from the crystal structure.

The weights are for the two pretrained Equivar models, BM1, and BM2. The included python scripts can be used for performing the regression of tensors of atomic Born effective charges using these models. The scripts can also be downloaded from https://github.com/equivar/equivar_eval/. Here, version v0.2.2 (a396545) is used.

The accompanying dataset was used for model training. The dataset contains tensors of Born effective charges of perovskite oxides, Li3PO4, and ZrO2, calculated from first principles using density functional perturbation theory (DFPT). The dataset is in the extended xyz format.